When to Use Oracle Database In-Memory

Transcription

1 When to Use Oracle Database In-Memory Identifying Use Cases for Application Acceleration O R A C L E W H I T E P A P E R M A R C H

2 Executive Overview Oracle Database In-Memory is an unprecedented breakthrough in Oracle database performance, offering incredible performance gains for a wide range of workloads. Oracle Database In-Memory can provide orders of magnitude performance improvements for analytics workloads, as well as substantial improvements for mixed-workload Enterprise OLTP applications. This document briefly introduces Database In-Memory, enumerates high-level use cases, and explains the scenarios under which it provides a performance benefit. The purpose of this document is to give you some general guidelines so that you can determine whether your use case is a good match for this exciting new technology. Introducing Database In-Memory Database In-Memory features a highly optimized In-Memory Column Store (IM column store) maintained alongside the existing row formatted buffer cache as depicted below in Figure 1. The primary purpose of the IM column store is to accelerate column-oriented data accesses made by analytics operations. It is similar in spirit to having a conventional index (for analytics) on every column in a table. However, it is much more lightweight than a conventional index, requiring no logging, or any writes to the database. Just as the performance benefit to an application from conventional indexes depends on the amount of time the application spends accessing data in the tables that are indexed, the benefit from the IM column store also depends on the amount of time the application spends on data access for analytic operations. It is therefore important to understand the basic characteristics of your application in order to determine the potential benefits from Database In-Memory. Figure 1: Dual format in-memory representation featuring new in-memory column store How Does Database In-Memory Improve Performance? The IM column store includes several optimizations for accelerated query processing. These are described in detail in the Database In-Memory Whitepaper so only a brief overview is provided here. There are four basic architectural elements of the column store that enable orders of magnitude faster analytic query processing: 1. Compressed columnar storage: Storing data contiguously in compressed column units allows an analytic query to scan only data within the required columns, instead of having to skip past unneeded data in other columns as would be needed for a row major format. Columnar storage therefore allows a query to perform 1 ORACLE DATABASE IN-MEMORY

3 highly efficient sequential memory references while compression allows the query to optimize its use of the available system (processor to memory) bandwidth. 2. Vector Processing: In addition to being able to process data sequentially, column organized storage also enables the use of vector processing. Modern CPUs feature highly parallel instructions known as SIMD or vector instructions (e.g. Intel AVX). These instructions can process multiple values in one instruction for instance, they allow multiple values to be compared with a given value (e.g. find sales with State = California ) in one instruction. Vector processing of compressed columnar data further multiplies the scan speed obtained via columnar storage, resulting in scan speeds exceeding tens of billions of rows per second, per CPU core. 3. In-Memory Storage Indexes: The IM column store for a given table is divided into units known as In-Memory Compression Units (IMCUs) that typically represent a large number of rows (typically several hundred thousand). Each IMCU automatically records the min and max values for the data within each column in the IMCU, as well as other summary information regarding the data. This metadata serves as an In-Memory Storage Index: For instance, it allows an entire IMCU to be skipped during a scan when it is known from the scan predicates that no matching value will be found within the IMCU. 4. In-Memory Optimized Joins and Reporting: As a result of massive increases in scan speeds, the Bloom Filter optimization (introduced earlier in Oracle Database 10g) can be commonly selected by the optimizer. With the Bloom Filter optimization, the scan of the outer (dimension) table generates a compact bloom filter which can then be used to greatly reduce the amount of data processed by the join from the scan of the inner (fact) table. Similarly, an optimization known as Vector Group By can be used to reduce a complex aggregation query on a typical star schema into a series of filtered scans against the dimension and fact tables. Apart from accelerating queries, Database In-Memory has the ability to speed up DML operations or writes to the database with the ability to replace analytic Indexes: Since the IM column store enables superfast analytics, it is possible to drop conventional indexes used only to accelerate analytic queries. Avoiding costly index maintenance allows update/insert/delete operations to be an order of magnitude faster. As stated earlier, the IM column store is a purely in-memory structure, and maintaining it is very low overhead. The following table summarizes the key application design principles for maximizing the benefits of Database In- Memory. None of these principles are new, but they are even more important to follow when using Database In- Memory because of the incredible speedup it provides for analytic data access. Table 1: General Guidelines for Maximizing the Benefits of Database In-Memory Rule 1 Rule 2 Process Data in the Database, not in the Application Process Data in Sets, not Row by Row Instead of reading rows out of the database into the application in order to compute a metric such as a total or an average, it is far more efficient to push that computation down into the database. This is especially true with Database In-Memory, since the benefits of processing within the database are much higher. This rule is applies to any analytics workload: The costs of database entry and exit are amortized by the number of rows processed. Since Database In-Memory can process billions of rows per second, it is important to give the database enough data to process on each invocation. 2 ORACLE DATABASE IN-MEMORY

4 Table 1: General Guidelines for Maximizing the Benefits of Database In-Memory Rule 3 Use Representative Optimizer statistics Plan differences can make a huge difference to the performance of a query especially when inmemory access paths can provide orders of magnitude faster performance. To ensure that you have optimal plans, please follow Oracle s recommended best practices for gathering a representative set of statistics. Rule 4 When Possible, Use Parallel SQL With Database In-Memory, IO bottlenecks are alleviated and CPU time dominates the overall execution profile. Parallelism is essential to maximize performance, using all available CPU cores for In-Memory processing. This is especially true in an Oracle Real Application Cluster environment, where Auto DOP is needed to fully utilize all available CPU cores. High-Level Use Cases for Database In-Memory As depicted in Figure 2 below, Database In-Memory can be used both within Enterprise OLTP systems and within Data Warehouses for real time analytics. Data Warehouse Systems For Data Warehouses, Database In-Memory can significantly improve the performance of analytics and reporting on data that can be accommodated within the IM column store, such as on table partitions representing relatively nearterm data. Tables within the Foundation layer and within the Access layer can leverage Database In-Memory. Due to the massive performance gains for queries on in-memory tables within the Foundation layer, it may be possible to eliminate many indexes and other summary objects such as (pre-computed cubes) from the Access layer. Database In-Memory is particularly applicable to data marts. Pre-computed summaries and aggregates (such as Key Performance Indicators), can usually be stored easily within memory. Note: Database In-Memory is generally not useful for the ETL or Staging layer where data tends to be written and read only once. Enterprise OLTP Systems Enterprise OLTP systems (those running packaged ERP, CRM, HCM applications such as Siebel, Peoplesoft, JD Edwards, etc) typically include a mixture of both OLTP transactions and periodic analytic reporting. In these systems, Database In-Memory can be used for real-time reporting against the base OLTP data. As stated earlier, the IM column store can potentially replace analytic indexes in these systems with significant speedups for OLTP DML operations. 3 ORACLE DATABASE IN-MEMORY

5 Removing analytic indexes results in many system-wide benefits, e.g. it reduces the total size of the database resulting in reduced storage requirements and faster backups. Analytic index removal also improves buffer cache hit rates, reduces overall redo and undo generation rates, and reduces the total I/O to the database. Figure 2: Use cases for Database In-Memory within the Enterprise Note: For specialized pure OLTP systems (such as real-time trading engines, real-time telecommunications billing or call routing applications) that do not have an Analytics component, there is no benefit from Database In-Memory. The Oracle TimesTen In-Memory Database is highly optimized for pure OLTP workloads providing response times in microseconds. For such systems, TimesTen may be a better choice if the data can be stored in-memory. Understanding Your Application Once there is a high-level match between Database In-Memory and your use case, it is important for you to understand where your application bottleneck is (i.e. where it spends the majority of its time) in order for you to estimate overall benefits from Database In-Memory. The abstract pie chart shown in Figure 3 below depicts a typical application time profile. Areas that Benefit from Database In-Memory As described earlier, Database In-Memory provides optimizations for dramatically faster Analytic queries. Therefore the following workload time components potentially benefit from Database In-Memory (as indicated in the pie chart): 1. Data Access for Analytics and Reporting: This is the core value proposition of Database In-Memory, to enable orders of magnitude faster analytic data access. 2. Analytic Index Maintenance: Database In-Memory often enables analytic indexes to be dropped, and eliminating the maintenance of these indexes improves overall application performance. Areas that do not benefit from Database In-Memory As the pie chart shows, there are also a number of other workload time components that do not benefit since they are unrelated to analytic data access or indeed, to any aspect of SQL execution. The same Oracle Database best practices that have been in vogue prior to the existence of Database In-Memory still apply when minimizing the time spent in these areas. 4 ORACLE DATABASE IN-MEMORY

6 Data Access - Analytics and Reporting * Application Network communication Database Logon/Logoff Parse PL/SQL execution SQL function execution Analytic Index Maintenance * Data Access - OLTP queries Data Writes (DML) to Database Tables OLTP Index Maintenance Figure 3: Abstract Time Profile for a Typical Application (* - time potentially reduced by Database In-Memory) 1. Application Time: Time within the application is unaffected by optimizations within the database. Application specific optimizations, including pushing more work into the database (as stated earlier in Table 1), are required to minimize this time. 2. Network communication between client and database: This is not affected by how fast the database runs. Standard techniques for reducing this time include using batched or array execution when possible to amortize the cost of database round-trips. 3. Logon and Logoff: Database connect / disconnect can be expensive, including the cost of authentication, process creation / teardown, etc. This time is completely unrelated to data processing. Standard techniques for minimizing this include keeping connections open and reusing connections when possible (e.g. by using connection pools). 4. Parse Time: While Database In-Memory can make the execution of SQL statements much faster, it has no impact on the time required to parse and optimize a SQL statement. Parse time should be minimized through common best practices by using bind variables, keeping cursors open, using session cached cursors, etc. The remaining areas that do not benefit from Database In-Memory are more fundamental to the application, and it is less likely that they can be addressed without some application redesign: 1. PL/SQL and SQL function execution: If the application spends most of its time within PL/SQL procedures, functions, or within built-in or user-defined SQL functions, then the bottleneck is in computation, and not in analytic data access. Database In-Memory does accelerate the evaluation of many common query predicates such as equality, range and list predicates, as well as many common aggregation functions such as MIN(), MAX(), SUM(), etc. In general however, a computation-intensive application will tend to benefit less from Database In-Memory. 5 ORACLE DATABASE IN-MEMORY

7 2. Data access by OLTP queries: OLTP queries, characterized by highly selective lookups (e.g. by Primary Key) or simple primary-foreign key joins, will not benefit from Database In-Memory. If a workload is dominated by this type of data access, the Oracle TimesTen In-Memory Database may be a better match (if the data being accessed can fit in memory). 3. OLTP index maintenance: Likewise, the indexes that are present in order to purely accelerate OLTP queries (such as Primary / Foreign key indexes) or those required for referential integrity, are still necessary even if Database In-Memory is used. Thus, the time spent in maintaining these indexes remains unchanged with Database In-Memory. Again, Oracle TimesTen may be a match for use cases dominated by this type of index access (if the data being accessed can be accommodated in memory). 4. Writes to Database Tables: The time spent on update/insert/delete DMLs against application tables will remain unchanged whether the tables are in memory or not. An application that is extremely write-intensive, and bottlenecked on DML on tables, will be unlikely to benefit significantly from Database In-Memory What Types of Queries Benefit from Database In-Memory So far we have explained some of the high-level use cases for Database In-Memory, and shown an abstract breakdown of application execution time to show what areas can benefit from Database In-Memory. It is also necessary to understand what types of analytic queries benefit most from the IM column store since not all queries will benefit equally. As a general rule of thumb: The greater the ratio of the total data accessed by a query to the data actually processed by the query, the greater the potential benefit from Database In-Memory. For example, let s consider the following scenario: 1) Query A scans a table with a million rows, but 990,000 rows are eliminated by the query predicates. 2) Query B scans the same million row table, but only 10,000 rows are eliminated by the query predicates. In this case, although both queries access 1 million rows, query A processes only 10,000 rows (since the rest are eliminated by the query predicates) while query B must process 990,000 rows (only 10,000 are eliminated). Query A therefore spends a far greater fraction of its total execution time on data access than Query B does. Query A will experience a greater reduction in execution time than Query B if the table is in-memory. To further illustrate this principle, we enumerate various query properties (properties impact how much a query can benefit from Database In-Memory) in Table 2 below. For each property, we show two queries that vary only in terms of the specified query property showing one query for which Database In-Memory provides a smaller benefit, and one for which it provides a larger benefit. This table is not intended to be a comprehensive list, rather to serve as a rough guide to help you understand what to look for when estimating the potential benefits of Database In-Memory. For these examples, we assume a simple star schema representing sales by an online marketplace: A single large SALES fact table and various much smaller dimension tables such as STORES, PRODUCTS, SHIPMENTS, CUSTOMERS, etc. 6 ORACLE DATABASE IN-MEMORY

8 Table 2:Typical Query Properties and how they impact the benefit from Database In-Memory Query Property Description Less Benefit Example Queries More Benefit Number of columns selected As more table columns are selected by a query, column processing costs start to dominate query execution time, reducing the benefit. SELECT * FROM Sales; Less benefit since the query selects all columns, and spends more time processing column values relative to the number of rows it accesses. SELECT revenue FROM Sales; More benefit since the query selects only 1 column and spends less time processing column values relative to the number of rows it accesses. Number of values returned The greater the number of values returned by a query, the smaller the IM benefit, because returning data back to the client will dominate the costs. SELECT revenue FROM Sales; Less benefit since the query returns a value for each row in the table. SELECT SUM(revenue) FROM Sales; More benefit since the query returns only one value even though it accesses all values in the column. Selectivity of column predicates A more selective column predicate enables more filtering on the scan results, and reduces the amount of data that intermediate query plan nodes need to process. Selective predicates can also leverage in-memory storage indexes to further accelerate data access. SELECT MEDIAN(revenue) FROM Sales WHERE revenue > 2; Less benefit since most rows will qualify (most sales are for items priced higher than $2) and a larger fraction of the query execution time will be spent in calculating the median. SELECT MEDIAN(revenue) FROM Sales WHERE revenue < 2; More benefit since fewer rows will qualify and a smaller fraction of the query execution time will be spent in calculating the median.. 7 ORACLE DATABASE IN-MEMORY

9 Table 2:Typical Query Properties and how they impact the benefit from Database In-Memory Query Property Description Less Benefit Example Queries More Benefit Selectivity of Join Conditions A more selective join condition will yield a smaller join result, and cause less data to have to be processed by the query. SELECT Si.d, S.revenue, P.name FROM Sales S,Products P WHERE S.prod_id=P.id; SELECT S.id, P.name, S.revenue FROM Sales S, Products P WHERE S.prod_id=P.id AND P.type= Footwear ; The above join will return a row for all sales records since each sale has a matching product. The query will spend far more time processing the join result This join will only return rows that correspond to sales of footwear products, a subset of total sales. As a result the query will spend much less time processing the join result. Number of tables being joined The greater the number of tables in a join, the larger the percentage of time spent in the join processing. SELECT <select list> FROM Sales, Products, Customers, Shipments, Stores, Suppliers, Warehouses WHERE <Join Condition> SELECT <select list> FROM Sales, Products, Customers WHERE <join condition> The above query involves a join between 7 tables and will spend more time in join processing as a fraction of the total execution time and will benefit less from in-memory. The above query involves a join between 3 tables and will spend less time in join processing as a fraction of total execution time, and will benefit more with the tables inmemory. Complexity of SQL functions Queries involving computationally expensive SQL functions will benefit less from in-memory. SELECT I.id, sum(s.revenue) FROM Sales S, Items I WHERE S.item_id = I.id AND MyMatch(I.name, LED TV )=1 GROUP BY I.id; SELECT I.id, sum(s.revenue) FROM Sales S, Items I WHERE S.item_id = I.id AND I.name LIKE %LED%TV GROUP BY I.id; The above query will spend more of in predicate evaluation using the user defined PL/SQL function MyMatch( ) and benefit less from Database In-Memory. The query will spend less time in predicate processing by applying the built-in fast string match LIKE operator to each item and benefit more from Database In-Memory. 8 ORACLE DATABASE IN-MEMORY

10 THE FACTS ABOUT ORACLE DATABASE IN-MEMORY Powering the Real-Time Enterprise Speed Up Analytics by Orders of Magnitude Oracle Database In-Memory transparently extends industry-leading Oracle Database 12c with columnar in-memory technology. Users get immediate answers to business questions that previously took hours because highly optimized in-memory column formats and SIMD vector processing enable analytics to run at a rate of billions of rows per second per CPU core. Unique Architecture Runs Analytics in Real-Time while Accelerating Mixed Workload OLTP Column format is optimal for analytics while row format is optimal for OLTP. Oracle Database In-Memory uses both formats simultaneously to allow real-time analytics on both Data Warehouses and OLTP databases. Indexes previously required for analytics can be dropped, accelerating mixed-workload OLTP. Compatible with All Existing Applications Industry-Leading Scale-Up Industry-Leading Scale-Out Industry-Leading High Availability and Security Cost Effective for Even the Largest Database Powering the Real-Time Enterprise Deploying Oracle Database In-Memory with any existing Oracle Database-compatible application is as easy as flipping a switch, no application changes are required. All of Oracle s extensive features, data types, and APIs continue to work transparently. Oracle s highly mature scale-up technologies enable application transparent In-Memory scale-up on SMP computers with up to tens of terabytes of memory and thousands of CPU threads. Data is analyzed at the enormous rate of hundreds of billions of rows per second with outstanding efficiency and no feature limitations. Oracle s highly mature scale-out technologies enable application transparent In-Memory scale-out across large clusters of computers with 100s of terabytes of memory and thousands of CPU threads. Data is analyzed at the enormous rate of trillions of rows per second with no feature limitations. Oracle s renowned Availability and Security technologies all work transparently with Oracle Database In-Memory ensuring extreme safety for mission critical applications. On Oracle Engineered Systems, In-Memory fault tolerance duplicates in-memory data across nodes enabling queries to instantly use an in-memory copy of data if a node fails. Oracle Database In Memory does not mandate that all data must fit in memory. Frequently accessed data can be kept In-Memory while less active data is kept on much lower cost flash and disk. The ability to easily perform real-time data analysis together with real-time transaction processing on all existing applications enables organizations to transform into Real-Time Enterprises that quickly make data-driven decisions, respond instantly to customer demands, and continuously optimize all key processes. 9 ORACLE DATABASE IN-MEMORY

11 Oracle Corporation, World Headquarters 500 Oracle Parkway Redwood Shores, CA 94065, USA Worldwide Inquiries Phone: Fax: Copyright 2014, Oracle and/or its affiliates. All rights reserved. This document is provided for information purposes only, and the contents hereof are subject to change without notice. This document is not warranted to be error-free, nor subject to any other warranties or conditions, whether expressed orally or implied in law, including implied warranties and conditions of merchantability or fitness for a particular purpose. We specifically disclaim any liability with respect to this document, and no contractual obligations are formed either directly or indirectly by this document. This document may not be reproduced or transmitted in any form or by any means, electronic or mechanical, for any purpose, without our prior written permission. Oracle and Java are registered trademarks of Oracle and/or its affiliates. Other names may be trademarks of their respective owners. Intel and Intel Xeon are trademarks or registered trademarks of Intel Corporation. All SPARC trademarks are used under license and are trademarks or registered trademarks of SPARC International, Inc. AMD, Opteron, the AMD logo, and the AMD Opteron logo are trademarks or registered trademarks of Advanced Micro Devices. UNIX is a registered trademark of The Open Group White Paper When To Use Oracle Database In-Memory March 2015 Author: Tirthankar Lahiri C O N N E C T W I T H U S blogs.oracle.com/in-memory